International Journal of Thermal Sciences 41 (2002) 11211129

www.elsevier.com/locate/ijts

Optimization of a buildings cooling plant

for operating cost and energy use
Darrell D. Massie
Department of Civil and Mechanical Engineering, United States Military Academy, West Point 10996-1792, USA
Received 20 February 2001; accepted 16 November 2001

Abstract
An optimal neural network-based controller for an ice thermal storage system has been developed and tested. The controller consists of
four neural networks, three of which map equipment behavior and one that acts as a global controller. The controller self-learns equipment
responses to the environment and then determines the control settings required to minimize operating cost. It has the advantage over other
controllers in that it always remains calibrated. Since it does not rely upon rules or assumptions, it is able to provide optimal control under any
utility pricing and operating condition. Although originally designed to minimize operating costs, simulation and optimization techniques
often determine minimum energy use as well.
2002 ditions scientifiques et mdicales Elsevier SAS. All rights reserved.

Keywords: Optimal control; Thermal energy storage; Chiller; Cooling; Least cost; Minimize energy; Neural network

1. Introduction higher ambient air temperatures. In general, rates are highest

Cooling of buildings in the United States is a major
contributor to the peak electrical load. By some estimates
it contributes up to 35 percent of the total electrical demand
in the United States, Henze [1]. As such, there is tremendous
potential to reduce operating cost and increase energy
efficiency with improved control. With the development of
advanced computer control techniques, costs can be lowered
without the need for trained technicians to continually
monitor equipment. This paper will address how neural
networks (NN) have been implemented in a laboratory
cooling plant to reduce cost and improve energy efficiency.
The chiller is a major Vapor Compression Refrigeration
Cycle (VCRC) component of a buildings cooling plant that
removes energy from chilled water, which is distributed to
cooling coils within the buildings mechanical system. Op-
erating costs associated with a buildings cooling plant are
often the highest in comparison with other mechanical com-
ponents (fans and pumps) within the building. VCRC equip-
ment is also generally more expensive to operate during the
day than at night due to time of use electricity pricing used
in many countries and higher condenser temperatures due to
E-mail address: id4747@usma.edu (D.D. Massie).
1290-0729/02/$ see front matter 2002 ditions scientifiques et mdicales Elsevier SAS. All rights reserved.
PII: S 1 2 9 0 - 0 7 2 9 ( 0 2 ) 0 1 3 9 8 - 4
during the day when electrical consumption is the greatest.
Using ice storage to cool commercial buildings, termed
thermal energy storage (TES), is a load management strategy
that can reduce electrical power or energy costs. Although
TES systems have been historically used so that smaller
chillers could be installed, most of todays installed thermal
storage systems are employed to shift the cost of electricity
from on-peak to off-peak periods, thus reducing demand
and energy charges. Unfortunately, many facility owners
are often disappointed with system performance since these
systems are not providing the expected load shifting. Poor
control has been identified as the primary reason for their
insufficient performance, Potter et al. [2]. Optimal control
of thermal storage with most of todays price structures is
difficult to establish because of the requirement to determine
building loads and equipment operating characteristics over
a planning horizon.
This paper describes a neural network controller that
learns the complex behavior of VCRC equipment and build-
ing loads and then optimally controls the system for least
cost. Neural networks, are a well-known tool among artifi-
cial intelligence techniques. They can reproduce the existing
relationship between input and output variables of complex
non-linear systems. Thus, they can be used to learn the be-
havior of complex cooling plants and then be used to control
1122 D.D. Massie / International Journal of Thermal Sciences 41 (2002) 11211129

Nomenclature

k hour of each day . . . . . . . . . . . . . . . . . . . . . . . 124 re energy charge at hour k of month . . . . $kWh1

J monthly power cost . . . . . . . . . . . . . . $month1 SCAP ice tank storage capacity . . . . . . . . . . . . . . kWhr
P (k) total power demand due to the cooling xk fraction of storage tank state of charge
and non-cooling load at hour k . . . . . . . . . . . kW at start of time k . . . . . . . . . . . . . . . . . . . . . . . . 01
Pmax, maximum power demand during period . . . kW xk+1 fraction of storage tank state of charge
Pnon-cooling non-cooling electrical load including of at the end of time k . . . . . . . . . . . . . . . . . . . 01
plug and lighting loads . . . . . . . . . . . . . . . . . . kW xmax maximum allowable charge state, set to 1.0.
Pplant cooling plant power demand including xmin minimum allowable charge state, set to 0.
compressors and the air-cooled condenser . kW
rd, demand rate incurred during hour of the
month . . . . . . . . . . . . . . . . . . . . . . . . . . . . . $kW1
them. Because the controller learns equipment operating
characteristics, equipment models change as equipment ages
or undergoes retrofit and there is no need to adjust the con-
troller once installed. It has advantages over other controllers
in that it provides optimal solutions and always remains cal-
ibrated.
2. Laboratory
The laboratory, Kreider et al. [3], where all equipment
testing was accomplished, is a full scale heating ventilating
and air conditioning (HVAC) laboratory, with a chiller
that is capable of satisfying 236 kW of cooling load
and representative of a typical floor of up to 930 m2
of a commercial building. The laboratory incorporates a
central hydronic heating and cooling plant, ice storage
tank, air-handling unit with variable frequency drives on
the fans, outside air conditioning station and four load
simulator zones, two of which are full scale. The laboratory
has a computer control and data acquisition system that
accommodates analog and digital inputs and outputs. The
chiller is a packaged dual-circuit unit with semi-hermetic
Fig. 1. Cooling plant configuration of Larson Laboratory.
t unit time step, one hour for this study . . . . . . hr
demand period . . . . . . . . . . . . . . . . . . . . . . . . . . . hr
number of days in the month . . . . . . . . . . . . 131
helical-rotary (screw) compressor. The ice storage tank is
a 6560 L, 2.4 GJ nominal capacity ice-on-coil with internal
melt storage system. Fig. 1 is a simplified schematic diagram
of the cooling plant.
Because the chiller plant is the primary focus of this
paper, it will be described in more detail. The 236 kW
chiller, uses R-22 refrigerant and has an overall full load
EER of 10.1 MBHkW1 . It is a packaged dual-circuit unit
with two semi-hermetic helical-rotary (screw) compressors,
a shell-and-tube evaporator and electronic expansion valves.
The unit comes with its own packaged controls that allow
low temperature operation for icemaking purposes, and
provides a low ambient lock-out and low water flow cut-
out. The typical operating temperature for the primary loop
is 3 C. Heat is rejected from the chiller through an air-
cooled condenser located outside the laboratory. The air-
cooled condenser was originally designed to serve a single
refrigerant circuit but has been retrofitted to serve two
independent and isolated circuits. It has a rated capacity, at
sea-level, of 176 kW. Each circuit is designed to turn on
and off with one of the compressors in the chiller. Each
circuit has three fans and three stages. The 176 kW rated
capacity is below the 236 kW capacity of the chiller at design
 D.D. Massie / International Journal of Thermal Sciences 41 (2002) 11211129
conditions, so unless the ambient temperature is below the
design temperature, the overall cooling plant capacity is
limited to 176 kW under design conditions.
The chiller plant is located upstream of a 665 kWh of
nominal capacity ice-on-coil storage packaged unit tank.
A three-way valve located in the primary loop controls the
amount of chilled water that flows through the ice storage
tank. A secondary loop allows chilled water to flow through
the air-handler that places the thermal load on the system.
The secondary loop is typically maintained at 7 C. There
are two constant speed pumps located in the primary chilled
water loop and one constant speed pump in the secondary
chilled water loop.
3. Reducing energy cost using thermal storage
Building owners are motivated to incorporate TES in a
cooling system to reduce operating costs. They are rarely
knowledgeable about the energy consumption rates of using
TES, which, usually consume more energy than direct
cooling systems due to the inefficiency of making ice
at subfreezing evaporator temperatures, Henze et al. [4],
Kintner-Meyer and Emery [5]. Subsequently, when there
is no economic incentive to freeze ice for later use, the
buildings cooling load can be met with direct cooling, which
generally consumes less energy.
A significant problem with controlling TES systems is
that ice must be formed in advance during periods of low en-
ergy cost so that the ice is available for cooling when energy
cost is high. This implies that accurate cooling plant equip-
ment models must be available to predict performance and
energy consumption over a wide variety of conditions. En-
ergy cost and equipment performance can then be estimated
over a planning horizon to determine if it is advantageous to
freeze ice in the near term to be used for later cooling.
Control strategies implemented in the field today do not
consider variations in building use and equipment operation
that can change from year to year, season to season or even
day to day. As a result, much of the potential cost savings of
using thermal storage systems is lost as cooling plants oper-
ate to meet conditions that are assumed and do not occur. For
example, many algorithms assume that a full charge of ice
might be needed during shoulder months, whereas this is of-
ten not the case. Sohn [6] also showed that equipment mod-
els that are developed by using manufacturers data or from
laboratory testing can vary significantly from field installed
equipment. Even though these difficulties have been recog-
nized, optimal control has not been implemented because
of complications accommodating the complex interactions
between equipment and the requirement for sensors. Equip-
ment behavior is highly non-linear and varies by location,
requiring experts to fine-tune and control. Even for experts
with vast experience in installing cool storage equipment,
models are complex and require significant effort to cali-
brate. Furthermore, as equipment ages or undergoes retrofit,
1123
models that describe equipment behavior must be changed,
requiring further expert assistance.
Several studies have addressed the need for improved
TES control. Akbari and Sezgen [7] point out that few TES
systems take advantage of daily variations in climate and
operating conditions so that charging and discharging are
optimized. This work also stresses the need for continued
research in TES systems. Different approaches have been
used to find optimal TES solutions. Braun [8] used an in-
dex of performance over a one-day period to minimize en-
ergy cost and Drees and Braun [9] developed a rule-based
controller to minimize cost without consideration of en-
ergy consumption. Simmonds [10] investigated energy con-
sumption but excluded the effect of price structure, which
is the primary concern of building owners. Kintner-Meyer
and Emery [11] investigated the sizing of thermal storage
components and their impact on the overall system cost and
in another study [12], investigated the use of an ice stor-
age facility in conjunction with the building thermal capaci-
tance. Henze et al. [13] developed a simulation environment
that used a realistic plant model to investigate the theoret-
ical limit of operating cost savings achieved by cool stor-
age. In another study, Henze and Krarti [14] showed that it
was possible to implement TES in such a way that operating
costs could be minimized while reducing energy consump-
tion. This work improves on previous work in that it incorpo-
rates price structure, equipment forecast modeling and cali-
bration, and an optimizer that determines what combination
of the above should be used to minimize cost.
3.1. Cooling plant control
This portion of the study describes the development of
an optimal controller that minimizes operating cost and
also minimizes energy usage. Equipment performance is
estimated using self-calibrating NN models developed by
Massie et al. [15]. Since these models self-calibrate to
installed operating performance, equipment modeling and
calibration problems are eliminated.
The chiller load is controlled by adjusting the cooling
plant temperature setpoint, which is the chilled water supply
temperature at the evaporator exit. The primary loop three-
way valve position determines the ice tank flow. The primary
loop three-way mixing valve located at the thermal storage
tank outlet determines how much of the water coming from
the chiller is circulated through the tank. When the valve
is set to 100% (termed 100% open), all fluid leaving the
chiller circulates through the ice tank. To charge the tank,
the chiller setpoint temperature must be below 0 C and the
valve must be opened. The lower the setpoint temperature
and more fully opened the valve, the faster the charge.
When discharging the tank, if the majority of cooling load
is to be met by the chiller, a combination of lowering
the chiller setpoint temperature (but still above freezing)
and closing the valve position will shift the load onto the
chiller. Likewise if more of the cooling load is to be met
1124 D.D. Massie / International Journal of Thermal Sciences 41 (2002) 11211129
by the ice tank, a combination of raising the chiller setpoint
temperature and opening the primary loop three-way valve
must be accomplished.
3.2. Planning horizon and cost function
In optimizing TES systems, the period of time (planning
horizon) that energy estimates are to be made into the
future must be selected. Henze et al. [16] found that a
planning horizon of 21 hours yielded cost differences of
less than two percent regardless of the predictor method
used. That study used a variety of estimators to include
perfect prediction, bin model, random walk harmonic model
and auto-regressive neural network. For the current study, a
conservative planning horizon of 24 hours was chosen due
to the cyclic nature of TES operation.
The objective function in a traditional cost function con-
sists of two parts, the cost of electrical energy [$kWh1 ]
consumed over the billing period and a cost for peak electri-
cal demand [$kW1 ]. For most US locations, electricity is
billed on a monthly basis using two distinct rate periods. The
cost J (expressed in monthly units) of operating the cooling
plant for one day can be simulated from




24

2 resented with a single variable that defines the fraction of
where Pnon-cooling is the non-cooling electrical load includ-
ing plug and lighting loads, Pplant is the power demand of
the cooling plant including compressors, the air-cooled con-
denser, all pumps in the primary and secondary loops and to
drive the variable speed fans of the air-handler units.
3.3. Simulation environment
The key feature of thermal storage is to minimize power
cost by bridging the temporary difference between cooling
load supply and demand. The power consumption, however,
is not a control variable, but is instead a result of operation
of the cooling system and non-cooling loads. In a system
without storage, the building load must be met immediately
by the chiller. With TES, the ice storage can be used to
meet the building load and the cost of replenishing the
storage moved to a period when electricity is less expensive.
Therefore, in a system with thermal storage, there is a choice
as to which source of cooling will be used at any particular
time. Cooling can be taken from the storage, the chiller, or
some combination of the two. This decision is based on a
comparison of operating costs.
The state of charge x of the storage tank can be rep-
J= P (k)re (k)t + Pmax, rd, [$month1] maximum ice formation. At any point in time, a decision is
1 k=1 =1 needed to either charge, discharge or leave the ice inventory
(1) unchanged. For ice storage systems the state transition equa-
tion is
where is the number of days in the month and k is the
t
hour of each day. P (k) is the total power demand due to xk+1 = xk + uk (4)
the cooling and non-cooling load at hour k, re is the energy SCAP
charge at hour k of month and t is the unit time step, subject to the constraints
which has been set to one hour in this study, although it xmin
xk+1
xmax (5)
could be any period of time. Demand charges are computed
by taking the product of the maximum power consumption where xk+1 is the state of charge of storage at the end of
Pmax, of the demand period and the demand rate rd, , that
is incurred during that hour of the month. For a utility tariff
that has two demand periods, would take on a value of 1
or 2.
An alternate price structure, real-time pricing (RTP), is
designed to charge the consumer more for electricity during
periods when electricity is more costly to produce. As a
result, there may or may not be a demand charge [$kW1 ]
as found in traditional rate structures. Indications are that
most true real-time pricing rate structures will consist of only
an energy charge [$kWh1 ] that will vary for each hour of
the day. Typically, the electricity supplier informs users of
next day rates in advance so that decisions on how to manage
electrical costs (to include ice storage) can be implemented.
The cost function for true RTP can then be written as




24
time k, t is the time step (one-hour), SCAP is the storage
capacity of the ice tank (e.g., kWh, BTU or ton-hours) and
uk is the charging (+) or discharging () rate of storage for
time step k. The minimum state of charge, xmin can be set to
zero if only the latent heat of fusion is to be considered or to
a negative value if sensible heat is to be used. The maximum
state of charge, xmax , is 1.0.
The rate of charge uk is automatically set by the NN
equipment models if the control settings for plant operation
are within the same range as those used to develop the
models. For example, if the range used for training a neural
network was between 4 C and 7 C, then the models
should not be applied outside of this range. Adjusting the
setpoint temperature at the chiller evaporator outlet controls
the chiller. A combination of chiller setpoint temperature and
the primary three-way valve determine the ice tank rate of
J= P (k)re (k)t month1 (2) charge uk .
1 k=1
3.4. Neural network controller architecture
Power demand P (k) includes the total of the cooling
plant and the non-cooling load and can be computed using Neural networks, are a well-known tool among artificial
P (k) = Pnon-cooling + Pplant (3) intelligence techniques. They can reproduce the existing
 D.D. Massie / International Journal of Thermal Sciences 41 (2002) 11211129
relationship between input and output variables of complex
non-linear systems. Thus, they can be used to learn the
behavior of complex cooling plants and then to control them.
Neural networks are particularly well suited for these types
of problems since they are easily configured to map several
input variables to multiple output variables. Neural networks
are trained to map an input vector to an output vector such
that error is minimized (e.g., Bishop [17] or Massie and
Curtiss [18]).
The setup of a neural network requires the choice of the
number of layers, the number of neurons in each layer, the
activation (transfer) function of each layer and the training
algorithm, Wasserman [19]. Two phases are then required
to make the neural network operative. The first phase is the
training (or learning) phase, in which the neural network
is taught to match a known set of corresponding input and
output values, in order to learn the relationship existing
between them. Training is achieved through the modification
of the weights associated with each neural connection.
This is done by the training algorithm, which aims to
minimize the error between predicted and actual values in
the training set. Training is the most time-consuming phase
and it is critical for the success of the neural network as a
predictive model. The second phase is called generalization
(or testing). The neural network is tested using another
known set of corresponding input and output values (none
of which belong to the training set) and its performance is
evaluated.
Neural networks offer the potential for control of proces-
ses through predictive techniques. Jordan and Rumelhart [20]
describe the construction of a composite learning system of
state-dependent mapping from inputs to predicted sensations
in a forward looking network. Anderson [21] described a
Fig. 2. Architecture of global controller.
1125
control algorithm where an inverted pendulum was balanced
using a neural network that controlled the movement of the
base. Curtiss [22] developed an algorithm in which neural
networks were used to control a heating coil and Curtiss et
al. [23] implemented a neural network-based energy man-
agement program that successfully performed on-line set-
point resets in an actual HVAC system without TES and
Jeannette et al. [24] developed a NN to control a heating coil
and boiler.
For this study, the neural network-based supervisory con-
troller used to determine hourly setpoints is a recurrent net-
work that computes output sequentially in time. Controlled
and uncontrolled input values are fed into the controller net-
work and the network modifies the weights associated with
the controlled input to minimize cost. For this network the
activation function is a combination of the NN equipment
models coupled with the hourly cost function.
The supervisory controller consists of two networks, a
training and predictor network, working in parallel (see
Fig. 2). The training network is used to learn the relationship
between the controlled and uncontrolled variables and the
plant characteristics, such as chiller power consumption
and tank charge/discharge rate. For example, the controlled
variables would be chiller setpoint temperature and ice tank
valve position over a 24 hour planning horizon. Examples
of uncontrolled variables would be outside air temperature,
utility tariffs and building cooling load. The training network
weights are then passed to the predictor network where they
are used in the activation function for the predictor network.
The predictor network subsequently finds values for the
control variables that minimize operating cost. Each network
operates independently depending on the need to find values
1126 D.D. Massie / International Journal of Thermal Sciences 41 (2002) 11211129
for control variables or the need to improve equipment and
forecasting models.
The predictor networks goal is to operate the cooling
plant for minimum cost. By fixing the equipment model
weights, the system is required to find an optimal solution
by varying only the control settings. Input for the controlled
settings come from virtual nodes whose activation is
always unity. Weights from the virtual nodes are trained
(i.e., modified) to attempt to reach an operating cost of zero.
A training goal of zero cannot be attained (unless the plant
is turned off), but a value of zero simplifies the mathematics
by removing it from the network and the back-propagated
error. If a linear activation function is used for the virtual
nodes, the weights become analogous to the setpoints. The
weights (setpoints) can also be constrained to meet any real-
world restrictions or desired limits.
3.4.1. Training network
The training network learns how a vector of controlled
and uncontrolled inputs will affect chiller and ice tank op-
eration. In short, this network trains the equipment mod-
els. Training is accomplished by collecting actual input and
output data over discrete time periods. The NN equipment
models are then compared to actual results and if the pre-
dicted output varies from actual equipment performance by
more than an allowable error, the data pair is included into
the training data set and the network retrained. The allow-
able error is set to 10% of equipment capacity. A 10% er-
ror tolerance was chosen based on findings of Drees and
Braun [25].
3.4.2. Predictor network
The structure of the predictor network is based on the
work of Nguyen and Widrow [26] and Curtiss et al. [27],
where each demonstrated how a recurrent network was
used to minimize a future error. The predictor network
receives weights from the training network and determines
the sequence of control actions that minimizes total cost over
a 24 hour horizon.
Calculations proceed by determining power consumption
and ice charge starting with the first hour and throughout the
planning window. Current conditions, at hour k, include in-
formation such as monthly peak demand and ice tank state
of charge. For each hour into the planning horizon, uncon-
trolled variables such as hourly ambient temperature and
anticipated building load are estimated. Hourly outside air
temperature can be estimated sufficiently well using Na-
tional Weather Service high and low temperature predictions
and the ASHRAE model discussed in Chapter 28, Table 2,
ASHRAE [28]. Building loads may be estimated by a variety
of methods, such as those listed in Kreider and Haberl [29].
The expected combined error of these estimates will vary by
building type and location, and will lead to a loss in opti-
mization accuracy of approximately 10%. Each hours cal-
culation estimates a cost and ice charge at the end of the
hour. If a control action attempts to use ice inventory that is
not available or fails to meet the buildings cooling load, a
cost penalty is applied and back propagated through the net-
work. The current hours final ice charge becomes the next
hours initial ice charge and so on. At the end of the plan-
ning window, hourly costs are summed and compared to the
desired cost of zero.
4. Results
The NN controller described here has been thoroughly
tested by computer simulation and on a full scale HVAC
system. Complete results that demonstrate the robustness
of this controller can be found in Massie [30] and Massie
and Bailey [31]. Provided here are examples to demonstrate
how both operating cost and energy consumption can be
reduced.
During the spring and fall seasons, cooling loads are typ-
ically smaller than during summer months. If the electri-
cal rate structure encourages load shifting, but the build-
ings cooling load is smaller than the storage capacity, then
only enough ice should be made during the off-peak pe-
riod to meet the building load for the next day. Freezing
more ice than required for the cooling load would incur ad-
ditional cost as well as energy penalties associated with re-
duction in efficiency while making ice at subfreezing evap-
orator temperatures. In this example, the on-peak demand
[$kW1 ] and energy [$kWh1 ] rates were set to five times
that of off-peak ratesa rate structure that strongly encour-
ages load shifting to off-peak periods. The on-peak period
runs from 0800 until 1700 daily and the buildings cooling
load is nearly flat and occurs during these same hours. There
is no cooling load during the off-peak period.
In this example, see Fig. 3, the buildings daily cooling
load was set to 60 percent of maximum ice storage capacity.
The controller recognized that with an on-peak to off-peak
cost ratio of five to one, that there was sufficient economic
incentive to move the cooling to the off-peak period and use
ice exclusively during the on-peak period. Additionally since
there is an off-peak demand charge, it was determined that
partial-loading the chiller at a nearly constant six to seven
kW was more cost effective than running the chiller at a
full load for a shorter period of time. This is the optimal
economic solution for the constraints given in this example
and also reduces energy consumption by only charging the
ice tank to 60 percent. There were some energy inefficiencies
associated with the part loading of the chiller, but they
were small in comparison to the economic gain. In general
screw type compressors have better part load efficiencies
than reciprocating compressors.
A comparison of these results will be made with the
simulation environment developed by Henze et al. [11]
and the rule based controller developed by Drees and
Braun [7]. These studies were chosen because they represent
the tremendous improvements in TES control that have been
made over the past decade. The optimal control environment
 D.D. Massie / International Journal of Thermal Sciences 41 (2002) 11211129 1127

Fig. 3. Shoulder month cooling load when a full charge of ice is not required.

Fig. 4. Consecutive days with load-shifting price incentive followed by no incentive.

of Henze, would have provided an optimal solution, if
equipment models had been accurately calibrated. For that
study, un-calibrated models were developed for the sole
purpose of comparing control strategies. However, since
between this study and Henzes, the ice tank and chiller
operated in a similar fashion, results are comparable. Results
of this study are improved over the Drees and Braun
solution. Their method assumes that a full charge is required
for the following day and ice is made unconditionally. Their
algorithm then discharges the tank so as to minimize cost.
Had their assumption of the requirement for a full charge
been correct, then their results would also be comparable.
Unfortunately, the penalty for making ice can be substantial.
Kintner-Meyer and Emery [10], for example, assumed a
39% performance penalty for making ice. Results of this
study show that the increased cost varies according to
environmental conditions and part-load ratios, however, a
penalty in excess of 30% was observed.
In the next example, Fig. 4, a two-day period is consid-
ered. For both days, as might be found during the height
of the cooling season, the buildings cooling load is much
greater than the storage capacity. The buildings cooling load
profile is unchanged for both days. A strong price incen-
tive (as in the previous example) is used during the first
day whereas there is no load shifting price incentive for the
second day. This could occur if the first day were a Friday
and the second a Saturday (assuming weekend rates are off-
peak).
The NN controller determined that ice should be made
for the first day and discharged during the period of high
costs. The electrical demand during the on-peak period
also remains flat to minimize cost. On the second day, the
controller uses direct cooling since there is no price incentive
to shift the cooling load and there is a thermodynamic
(and economic) penalty if ice were frozen to meet the
second days load. This demonstrates the controllers ability
1128 D.D. Massie / International Journal of Thermal Sciences 41 (2002) 11211129
to respond to price signals. As a side note, since the
planning horizon for the NN controller is 24 hours, the
controller considers each planning window as a separate
event. However, because of the cyclic nature of TES use,
the solution is optimal.
With real-time pricing (RTP), conventional control strate-
gies become even less effective since they are based on as-
sumptions that will not apply as rates vary throughout the
day. Solutions under RTP are more difficult to verify be-
cause economics and equipment performance vary hourly,
if not more frequently. To demonstrate the robustness of the
NN controller, the RTP rate shown in Fig. 5 was used. As-
suming utility tariffs are known 24 hours in advance, there
is a three-hour period (from 1100 to 1300) where the utility
rates are highest and use of ice storage should be maximized.
We also note that the next highest tariff is found at 1400 and
so an optimal solution would maximize use of the TES next.
The building load profile is the same as that used for the last
example.
Fig. 5. RTP price structure.
Fig. 6. Consecutive days with load-shifting price incentive followed by no incentive.
The optimal solution is shown in Fig. 6. The solution
shows that the optimal trajectory is to build ice inventory
during the night when rates are low. During the three hours
when rates are highest, the chiller is turned off completely
and all cooling comes from the ice tank. Between low
cost and peak cost hours the chiller is operated such that
the building load is met and storage is for the most part
neither increased nor decreased. The exception to this is
at 1400 (second most costly per kilowatt-hour), when the
remaining ice inventory, not used during the peak period, is
depleted. These results agree with expected results for this
pricing.
5. Conclusions
A neural network-based optimal controller has been de-
veloped to control a commercial ice storage system for least
cost. It is predictive and considers building load forecast-
ing, equipment behavior and utility rate structures. The con-
troller uses dual networks, one as a training network that
is used to create and train processes and a second that cas-
cades the processes developed by the training network to de-
termine setpoints over a planning window, here, 24 hours.
Although designed to operate for least cost, it will of-
ten operate using minimal energy as well. Since the con-
troller does not rely on assumptions, it is robust in find-
ing solutions given any price structure, building cooling
load and equipment operating conditions. Because of its
ability to learn patterns, it self calibrates to equipment op-
erating characteristics and does not require an expert to
fine tune. This feature insures that the controller will oper-
ate optimally as a building or equipment undergoes retro-
fit.
 D.D. Massie / International Journal of Thermal Sciences 41 (2002) 11211129
References
[1] G.J. Henze, Evaluation of optimal control for ice storage systems,
Ph.D. Dissertation, University of Colorado, Boulder, CO: Department
of Civil, Environmental and Architectural Engineering, 1995.
[2] R.A. Potter, D.P. Weitzel, D.J. King, D.D. Boettner, ASHRAE RP-
766: Study of operational experience with thermal storage systems,
ASHRAE Trans. 101 (2) (1995) 549557.
[3] J.F. Kreider, P.S. Curtiss, D.D. Massie, E. Jeannette, A commercial
scale University HVAC laboratory, ASHRAE Trans. 105 (1) (1999).
[4] Development of a predictive optimal controller for thermal energy
storage systems, Internat. J. Heating, Ventilating, Air-Conditioning
and Refrigeration Research 3 (3) (1997) 233264.
[5] M. Kintner-Meyer, A.F. Emery, Optimal control of an HVAC system
using cold storage and building thermal capacitance, Energy and
Buildings 23 (1995) 1931.
[6] C.W. Sohn, Field performance of an ice harvester storage cooling
system, ASHRAE Trans. 97 (2) (1991) 11871193.
[7] H. Akbari, O. Sezgen, Case studies of thermal energy storage (TES)
systems, evaluation and verification of system performance, LBL-
30852, Lawrence Berkeley Laboratory, University of California:
Energy and Environment Division, 1992.
[8] J.E. Braun, A comparison of chiller-priority, storage-priority, and
optimal control of an ice storage system, ASHRAE Trans. 98 (1)
(1992) 893902.
[9] K.H. Drees, J.E. Braun, Development and evaluation of a rule-
based control strategy for ice storage systems, Internat. J. Heating,
Ventilating, Air-Conditioning and Refrigeration Research 2 (4) (1996)
312336.
[10] P. Simmonds, A comparison of energy consumption for storage
priority and chiller priority for ice-based thermal storage systems,
ASHRAE Trans. 100 (1) (1994) 17461753.
[11] M. Kintner-Meyer, A.F. Emery, Cost optimal analysis and load
shifting potentials of cold storage equipment, ASHRAE Trans. 101 (2)
(1995) 539548.
[12] M. Kintner-Meyer, A.F. Emery, Optimal control of an HVAC system
using cold storage and building thermal capacitance, Energy and
Buildings 23 (1995) 1931.
[13] G.J. Henze, M. Krarti, M.J. Brandemuehl, A simulation environment
for the analysis of ice storage controls, Internat. J. Heating, Ventilating,
Air-Conditioning and Refrigeration Research 3 (2) (1997) 128148.
[14] G.J. Henze, M. Krarti, Ice storage system controls for the reduction of
operating cost and energy, in: Proceedings of the ASME International
Solar Energy Conference, American Society of Mechanical Engineers,
1996.
[15] D.D. Massie, P.S. Curtiss, J.F. Kreider, Predicting central plant HVAC
equipment performance using neural networksLaboratory System
Test Results, ASHRAE Trans. 104 (2) (1998) 221228.
[16] G.J. Henze, R.H. Dodier, M. Krarti, Development of a predictive opti-
mal controller for thermal energy storage systems, Internat. J. Heating,
1129
Ventilating, Air-Conditioning and Refrigeration Research 3 (3) (1997)
233264.
[17] C.M. Bishop, Neural Networks for Pattern Recognition, Clarendon
Press, Oxford, 1995.
[18] D.D. Massie, P.S. Curtiss, Neural network basics for use in building
mechanical systems, in: Proceedings from 2001 International Con-
ference of Chartered Institution of Building Services Engineers and
American Society of Heating Refrigeration and Air-Conditioning En-
gineers, Inc., 2001.
[19] P.D. Wasserman, Neural Computing: Theory and Practice, Van Nos-
trand Reinhold, New York, 1989.
[20] M.I. Jordan, D.E. Rumelhart, Forward models: Supervised learning
with a distal teacher, Cognitive Sci. 16 (1992) 307354.
[21] C.W. Anderson, Learning to control an inverted pendulum using neural
networks, IEEE Control Systems Magazine (April 1989) 3136.
[22] P.S. Curtiss, Experimental results from a network-assisted PID con-
troller, ASHRAE Trans. 102 (1) (1996) 11571168.
[23] P.S. Curtiss, M.J. Brandemuehl, J.F. Kreider, Energy management in
central HVAC plants using neural networks, ASHRAE Trans. 100 (1)
(1994) 476493.
[24] E. Jeannette, K. Assawamartbunlue, P.S. Curtiss, J.F. Kreider, Ex-
perimental results of a predictive neural network HVAC controller,
ASHRAE Trans. 104 (2) (1998).
[25] K.H. Drees, Braun, Modeling of area-constrained ice storage tanks,
Internat. J. Heating, Ventilating, Air-Conditioning and Refrigeration
Research 1 (2) (1995) 143159.
[26] D.H. Nguyen, B. Widrow, The truck backer-upper: An example of
self-learning in neural networks, in: Proceedings of the International
Joint Conference on Neural Networks, 1989, pp. 357363.
[27] P.S. Curtiss, J.F. Kreider, M.J. Brandemuehl, Adaptive control
of HVAC processes using predictive neural networks, ASHRAE
Trans. 99 (1) (1993) 496504.
[28] ASHRAE, ASHRAE Fundamentals Handbook, American Society of
Heating, Refrigerating and Air-Conditioning Engineers, Atlanta, GA,
1997.
[29] J.F. Kreider, J.S. Haberl, Predicting hourly building energy use: The
great energy predictor shootout-overview and discussion of results,
ASHRAE Trans. 100 (2) (1994) 11041118.
[30] D.D. Massie, Neural network controller for commercial ice thermal
storage systems, Ph.D. Dissertation, University of Colorado, Boulder,
CO: Department of Civil, Environmental and Architectural Engineer-
ing, 1998.
[31] D.D. Massie, M.B. Bailey, Optimization and fault evaluation of a
buildings cooling plant using neural networks, in: ECOS 2000, In-
ternational Conference on Efficiency, Cost, Optimisation, Simulation
and Environmental Aspects of Energy and Process Systems, Univer-
sity of Twente, Enschede, The Netherlands, July 57, ASME, 2000,
pp. 941954.